这里会有这个项目的代码详解和我的一些ocr经验和心得,我会慢慢更新,有兴趣可以看看,希望可以帮到新接触ocr的童鞋CSDN博客
最近跟新:
- 2021.03.26 更新CRNN 训练效果,代码整理后上传
- 2021.03.06 更新CRNN backbone resnet 和 mobilev3 以及配置文件
- 2020.12.22 更新CRNN+CTCLoss+CenterLoss训练
- 2020.09.18 更新文本检测说明文档
- 2020.09.12 更新DB,pse,pan,sast,crnn训练测试代码和预训练模型
目前已完成:
接下来计划:
- 模型转onnx及调用测试
- 模型压缩(剪枝)
- 模型压缩(量化)
- 模型蒸馏
- tensorrt部署
- 训练通用化ocr模型
- 结合chinese_lite进行部署
- 手机端部署
使用 MJSynth(MJ) 和 SynthText(ST) 训练,以batchsize=512训练,在以下数据集上测试:
模型 | 迭代次数 | CUTE80 | IC03_867 | IC13_1015 | IC13_857 | IC15_1811 | IC15_2077 | IIIT5k_3000 | SVT | SVTP | mean |
---|---|---|---|---|---|---|---|---|---|---|---|
resnet34+lstm+ctc | 120000 | 82.98 | 91.92 | 90.93 | 91.59 | 73.10 | 67.98 | 90.16 | 85.16 | 78.29 | 83.56 |
mobilev3_large+lstm+ctc | 210000 | 73.61 | 92.50 | 90.34 | 91.59 | 74.82 | 68.89 | 87.56 | 83.46 | 77.20 | 82.21 |
mobilev3_small+lstm+ctc | 210000 | 66.31 | 90.77 | 88.76 | 91.13 | 73.66 | 69.52 | 88.80 | 84.54 | 72.24 | 80.64 |
训练只在ICDAR2015文本检测公开数据集上,算法效果如下:
模型 | 骨干网络 | precision | recall | Hmean | 下载链接 |
---|---|---|---|---|---|
DB | ResNet50_7*7 | 85.88% | 79.10% | 82.35% | 下载链接(code:fxw6) |
DB | ResNet50_3*3 | 86.51% | 80.59% | 83.44% | 下载链接(code:fxw6) |
DB | MobileNetV3 | 82.89% | 75.83% | 79.20% | 下载链接(code:fxw6) |
SAST | ResNet50_7*7 | 85.72% | 78.38% | 81.89% | 下载链接(code:fxw6) |
SAST | ResNet50_3*3 | 86.67% | 76.74% | 81.40% | 下载链接(code:fxw6) |
PSE | ResNet50_7*7 | 84.10% | 80.01% | 82.01% | 下载链接(code:fxw6) |
PSE | ResNet50_3*3 | 82.56% | 78.91% | 80.69% | 下载链接(code:fxw6) |
PAN | ResNet18_7*7 | 81.80% | 77.08% | 79.37% | 下载链接(code:fxw6) |
PAN | ResNet18_3*3 | 83.78% | 75.15% | 79.23% | 下载链接(code:fxw6) |
这里使用mobilev3作为backbone,在icdar2015上测试结果,未压缩模型初始大小为2.4M.
- 对backbone进行压缩
模型 | pruned method | ratio | model size(M) | precision | recall | Hmean |
---|---|---|---|---|---|---|
DB | no | 0 | 2.4 | 84.04% | 75.34% | 79.46% |
DB | backbone | 0.5 | 1.9 | 83.74% | 73.18% | 78.10% |
DB | backbone | 0.6 | 1.58 | 84.46% | 69.90% | 76.50% |
- 对整个模型进行压缩
模型 | pruned method | ratio | model size(M) | precision | recall | Hmean |
---|---|---|---|---|---|---|
DB | no | 0 | 2.4 | 85.70% | 74.77% | 79.86% |
DB | total | 0.6 | 1.42 | 82.97% | 75.10% | 78.84% |
DB | total | 0.65 | 1.15 | 85.14% | 72.84% | 78.51% |
模型 | teacher | student | model size(M) | precision | recall | Hmean | improve(%) |
---|---|---|---|---|---|---|---|
DB | no | mobilev3 | 2.4 | 85.70% | 74.77% | 79.86% | - |
DB | resnet50 | mobilev3 | 2.4 | 86.37% | 77.22% | 81.54% | 1.68 |
DB | no | mobilev3 | 1.42 | 82.97% | 75.10% | 78.84% | - |
DB | resnet50 | mobilev3 | 1.42 | 85.88% | 76.16% | 80.73% | 1.89 |
DB | no | mobilev3 | 1.15 | 85.14% | 72.84% | 78.51% | - |
DB | resnet50 | mobilev3 | 1.15 | 85.60% | 74.72% | 79.79% | 1.28 |
微信号:-fxwispig-
- https://github.com/PaddlePaddle/PaddleOCR
- https://github.com/whai362/PSENet
- https://github.com/whai362/pan_pp.pytorch
- https://github.com/WenmuZhou/PAN.pytorch
- https://github.com/xiaolai-sqlai/mobilenetv3
- https://github.com/BADBADBADBOY/DBnet-lite.pytorch
- https://github.com/BADBADBADBOY/Psenet_v2
- https://github.com/BADBADBADBOY/pse-lite.pytorch